Karpathy shows LLM can convincingly argue both sides of the same case
Andrej Karpathy described using an LLM to sharpen a blog post, then watched it dismantle its own argument when asked to take the opposite side.
Andrej Karpathy posted on X that he spent four hours using an LLM to meticulously refine the argument in a draft blog post, finishing convinced the piece was persuasive. On a whim, he then asked the same model to argue the opposite position — and, in his telling, it demolished the original argument and left him convinced the reverse was true.
The anecdote lands on a well-known property of current language models: they optimize for rhetorical coherence rather than truth, and will produce equally fluent cases for contradictory claims on demand. For people who build, buy, or use AI, it is a reminder that using LLMs as writing or reasoning partners can amplify confidence without improving correctness. Teams relying on these tools for analysis, policy drafting, or decision support increasingly pair them with adversarial prompts, red-team passes, or human review to surface the counter-case before shipping.